Developing AI models or giant GPU clusters? Uncle Sam would like a word
theregister.com
theregister.com
Outside of reporting that you're working on a new huge model and some maybe down the road future NIST guidelines they don't actually restrict the production of models at all. It's all about telling the humans jumping on the move fast break things train that no you can't use these tools for for <obviously dystopian thing> like renter screening, price fixing, and judicial sentencing.
Good on the white house for recognizing that the harm these models produce is almost entirely the humans connecting them to the real world.
I’d prefer to see strategic action on concrete issues rather than perceived risks. I don’t see how tasking a bunch of government agencies with pontificating on new and unfamiliar technologies can produce a good result.
I get the don't give an inch or they will take a mile, but also some regulation structures are probably warranted by such a fundamental shift on par with the advent of the Internet.
Back then we both got it right with DMCA in that big Internet companies could thrive in the US; but also problematic concentration of Monopolistic power in big Internet companies.
Will have to do some iteration to see what makes sense with the advent with this new computing model / capabilities at this scale.
It's fine to say you can't legally use software to commit crimes.
On an unrelated note, OpenAI announces GPT5 will be trained with fixed-point arithmetic.
https://twitter.com/DavidVorick/status/1719097248699879831
> (i) any model that was trained using a quantity of computing pow greater than 10^26 integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 10^23 integer or floating-point operations;
You can estimate it quite accurately, actually
You could probably get a better estimate based on the network structure, the number of training epochs, and the size of each batch.
And history has shown that companies don't just move to China/Russia because the US market is so lucrative.
And their phones are manufactured in China, India and Vietnam.
Nobody said that. Please don't do straw men arguments.
They moved most of their manufacturing to China.
2. The export controls have exceptions, loopholes, and people with strong wills.[1][2]
3. China is slowly but steadily freeing itself from import dependence.[3]
[1]: https://tech.slashdot.org/story/23/08/21/203241/china-keeps-...
[2]: https://mobile.slashdot.org/story/23/09/09/1849221/huawei-sh...
[3]: https://mobile.slashdot.org/comments.pl?sid=23071327&cid=638...
There is an argument that China has the world's largest economy and it is nestled in the middle of the region that is best known for high-tech manufacturing. It is still possible that they'll fumble this somehow, but the fundamentals are solidly on the side of China becoming the place to do AI training.
It’s not as secure and stable of a situation as you present.
It then sets every federal agency out on a quest to identify and then create a plan to ameliorate all the "scary AI boogey men" sci-fi fever dreams that have been associated with the deployment of a technology that doesn't even actually exist yet.
And, of course, more H1-B visas, because.. you know.. we wouldn't want to be "left behind."
[0]: https://www.whitehouse.gov/briefing-room/presidential-action...
The H1-B thing is kind of a non-sequitur. I think anyone involved in tech (software, hardware, anything) should be wildly in favor of massive H1-B increases, AI-related or not.
The U.S. has the opportunity to cement itself as the center of the world as far as technological progress from now into eternity. Why wouldn’t we take every brilliant immigrant we can get?
If you’re a software developer and you are afraid for your job/wages, I get that intuitively; but how many times do we have to learn the lesson that “more people making software leads to more software jobs”
Not if you like high wages, and you want them to stay that way.
The danger, IMHO, isn't wage dilution anyway (that one can be counteracted by politics) - it is that sensitive knowledge will make its way back to China and other current or potentially hostile nations, and it would not be the first time either that this happens.
Devs have been worried about offshoring/immigrants replacing them/lowering their wages for decades.
The only outcome we have ever observed from having more people building software is that dramatically more software jobs have become possible and in-demand.
Also, and I understand why you might not make this argument, but let’s be honest: software engineering wages are extraordinarily high. Slowing or even reversing that growth by small amounts on average would still leave millions of extremely well-paid jobs.
Software is eating the world after all. Seems likely that demand would be high anyway. Perhaps if there were no H1Bs, entry level grads would make $250k, with the average senior dev making $750k+.
YMMV, I guess.
I don't accept this premise and I don't think serves as a reasonable excuse for government interference in labor markets. Even if you do accept this, then the solution sacrifices long-term labor stability for short-term labor monopolization.
Either way, I don't see this as a positive outcome, and I regret every administrations attempt to expand the program using any excuse that happens across their desks.
> but how many times do we have to learn the lesson that “more people making software leads to more software jobs”
The connection between this outcome and increased H1-B visas for mostly _corporate sponsors_ is sketchy, at best.
b) There is no evidence that H1B visas have caused labor instability in the IT market.
Lmao. It’s self evident that the IT market, exemplified by the wealthiest tech corps, is literally dominated by foreign workers. The argument for unskilled labor (Americans don’t want those jobs) can’t even be dishonestly argued here. It’s government policy allowing US workers to be sidelined in favor of foreign ones.
In my view it's the opposite, a perfectly free labor market would be one where anyone can apply to a job. Restricting immigration by denying visas is a government interference in the market. So more visas means less interference.
(Note: I'm not claiming anything about whether it's a good idea or not).
The US doesn’t need to “cement itself” as anything. Thanks, but no thanks. Ironically, your kind of attitude is the same one that gave rise to the insane, populist politics of the last few years and has done more harm to immigrants than any other single policy.
H1-B is specifically for specialized occupations and generally requires a minimum of a bachelor’s degree and specialized skills in demand. It is, in fact, called the H1-B Specialty Occupations Visa.
I suspect you’re thinking of the Diversity Visa program, which offers a lottery of 55,000 visas annually to anyone (except for some eye brow raising exceptions).
There are also other programs like migrant worker programs that allow unskilled labor into the country for a limited time to fill seasonal work gaps.
> those workers would qualify for skilled worker visas.
H1B is a skilled worker visa so I’m not sure what the complaint here is.
It sounds like a lot of complaints you have are with well documented abuses of the program by consulting firms and the government’s lack of cracking down on that.
Luckily you don’t control the thoughts of others. US companies should hire US citizens and if there isn’t enough supply we should be enacting policies to fix that issue not continuing to allow US citizens to work “unskilled” labor for pennies while we brain drain other nations and create an upper class of non-citizens.
b) It is inarguable that we are in an age of AI.
c) There are legitimate concerns with AI that shouldn't be blindly waved away as "boogey man". Especially with what we have been seeing in Ukraine with their use of autonomous and remotely controlled drones. Adding an AI layer into the mix needs to be regulated.
d) Number of H1B visas are set by Congress not the President.
How many days from today until that election day?
> It is inarguable that we are in an age of AI.
We have companies hawking large language models. It's entirely arguable that we are in an "age of AI." You're about to be in an age of "no more easy CMOS gains." The intersection between these two points is going to be interesting.
> There are legitimate concerns with AI that shouldn't be blindly waved away as "boogey man"
This is why we have courts and a legislature with committee powers. I do not believe that an eager top down federal agency approach is going to solve real problems without creating more encumbrances than it's worth.
This is also why it's pandering. Look at the list of issues they bring up, those most definitely rank with voters.
> Number of H1B visas are set by Congress not the President.
Well then it's potentially even a bigger problem. They're going to change prioritization for those applicants against a limited pool.
I know there is some work being done to 'extract' training data from models (although given the compression, not entirely sure you could really extract everything but curious if anyone is seriously working on training models on purely encrypted data
You have been sorely misinformed about AI. Even the name itself has been used to mislead you! Artificial Intelligence does not exist today. While new systems may be "intelligent" in their designs, none of them is "an intelligence".
Artificial Intelligence has been the pursuit of Computer Science since the earliest days of software design, back when the AI department at Bell Labs developed Programming Languages. What is called "AI" today is no more than the newest efforts in that pursuit. Despite the excitement at these newer efforts, the goal that is AI is as mysterious as it ever was.
So what's new? Inference Models. These allow computers to navigate ambiguous data. This was entirely impossible before, and is only somewhat possible today. While computers do not get completely stuck on ambiguity the way they used to, they are still unable to conclusively resolve that ambiguity. They can only be trained to guess. With careful training on very large datasets, some impressive results have been obtained. Unfortunately, those impressive results are always closely tied to embarrassing mistakes. This is a feature of contemporary inference models.
The goal in mind is to perform this process well enough to have a novel and useful system. Many believe that once a model is big enough and trained well enough that it will become reliably more accurate. There is no conclusive evidence that is the case. While they are often introduced as a "limitation", behaviors like "overconfidence" and "hallucinations" are features of these systems. In order to remove a feature from a system, a new system must be invented. In the mean time, let's consider what does exist, rather than get lost in our own dreams.
So what should you be worried about? Contemporary inference models are powerful enough to create convincing results. What does that mean for the people of the United States? We need to recognize the reality in front of us: people can tell lies. Data alone is never a reliable source of information. This has always been true, but the inherent difficulty in storytelling has made some lies impractical to tell. As technology improves, so does each person's ability to tell a story.
We have all watched the journalistic integrity of our world suffer at the hands of Social Media, and at the failure of large corporations to moderate content. At best, people have grown to distrust scientific discovery and leadership; and at worst, untamed hate speech has lead to genocide.
So what can we do about it? Readers need to be able to differentiate content, not by its substance, but by its source. The good news is that this has been a solved problem for 50 years. All that an author needs to attach their identity to their writing, is to provide readers their unique public key and a signature of their work. Unfortunately, the best tools to do this, including GPG, are very technical and difficult for the layman to use. It should be the priority of the United States, for the sake of national security, to improve this landscape by creating (or motivating the creation of) easy-to-use public-key encryption software.
What is it that you are saying does not exist?
Is that the same thing as you think policy makers or whatnot have been led to believe exists?
Artificial intelligence is intelligence created by humans. Artificial means man-made.
I had thought the fact that this is what I meant, would have been clear in my initial reply. I didn’t expect this conversation to go in the direction of clarifying that.
(*): modulo negligible quibbles about wording
The technologies we have are models, not actors. They are each the static result of a process that determines boundaries and relationships between pieces of data. These models do not, however, organize or label the boundaries or relationships themselves.
For example, you can model a dataset of human-written text into an LLM. Using that LLM, you can transform a human-written prompt into a "continuation" that incorporates the modeled dataset. The resulting continuation will contain a new organization of both text from your prompt and/or text from the dataset: nothing more.
The boundaries and relationships modeled by an LLM are not categorized. The model does not contain any objective observation about its data. It only provides a structure that is intended to "align to" (simulate) the already-present semantics of natural language. It was not the model's intention to align with language semantics: that comes from the authors of the model, and from the presence of patterns (we recognize as language semantics) in the dataset itself. Without the presence of natural language patterns, there would be no boundary to align to in the first place.
To contrast, a human can read text into ideas, then think objectively about those ideas, produce new ideas, and finally express those ideas into text by structuring them into language semantics. Nothing like that happens in any software I am aware of.
First, the term backdoor applies loosely to the Clipper chip since it would have been public that this chip could be used for accessing private information. I think lock is a better security term. Backdoors generally are secret. My memories from that time are that the term backdoor was also an irony.
Second, the link is about informing private information because there is a security concern. It is not because the government want to have stats to inform the public about how to run AI nodes.
As long as we're nitpicking word choice, the word allegory does not apply here. An allegory is an intentional narrative device employed by an author or artist. Two things you find to be similar does not constitute an allegory.
The link you identify is pretty tenuous. With the Clipper chip, the government wanted access to live communications, not information about the infrastructure the communications were passing through. That is not the case with the recent EO.
Your oath of office said “I do solemnly swear (or affirm) that I will support and defend the Constitution of the United States against all enemies, foreign and domestic; that I will bear true faith and allegiance to the same; that I take this obligation freely, without any mental reservation or purpose of evasion; and that I will well and faithfully discharge the duties of the office on which I am about to enter: So help me God.” Source: https://oaths.us/senate-oath-of-office/
I noticed that you just voted on a bunch of unconstitutional laws, which makes you guilty of perjury.
Note: It does not matter which memeber of Congress you address, or when. This applies to all of them, all the time.